Perspectives on model forecasts of the 2014-2015 Ebola epidemic in West Africa: lessons and the way forward

Abstract

The unprecedented impact and modeling efforts associated with the 2014-2015 Ebola epidemic in West Africa provides a unique opportunity to document the performances and caveats of forecasting approaches used in near-real time for generating evidence and to guide policy. A number of international academic groups have developed and parameterized mathematical models of disease spread to forecast the trajectory of the outbreak. These modeling efforts often relied on limited epidemiological data to derive key transmission and severity parameters, which are needed to calibrate mechanistic models. Here, we provide a perspective on some of the challenges and lessons drawn from these efforts, focusing on (1) data availability and accuracy of early forecasts; (2) the ability of different models to capture the profile of early growth dynamics in local outbreaks and the importance of reactive behavior changes and case clustering; (3) challenges in forecasting the long-term epidemic impact very early in the outbreak; and (4) ways to move forward. We conclude that rapid availability of aggregated population-level data and detailed information on a subset of transmission chains is crucial to characterize transmission patterns, while ensemble-forecasting approaches could limit the uncertainty of any individual model. We believe that coordinated forecasting efforts, combined with rapid dissemination of disease predictions and underlying epidemiological data in shared online platforms, will be critical in optimizing the response to current and future infectious disease emergencies.

Keywords: Data sharing; Disease forecast; Ebola; Epidemic model; Exponential growth; Lessons learned; Polynomial growth; Sub-exponential growth; West Africa.

Fig. 1

Observed trajectory of the Ebola epidemic in the three most affected countries of West Africa against predictions made in the midst of the outbreak. The colored horizontal lines represent model predictions for Guinea (G), Liberia (L), Sierra Leone (S), or all three countries combined (WF); the beginning of the line is when the prediction was made, whereas the end of the line marks the date the prediction is for (thus, shorter horizontal lines illustrate near-term predictions, while longer lines illustrate further time horizons). Data taken from JP Chretien’s Elife review [4]

Fig. 2

Google search volume – a signal that quantifies people’s web searches – for the phrase “Exponential Ebola” quickly surged and became popular during weeks 30–40, coinciding with the epidemic growth of reported cases in West Africa (Spearman’s rho = 0.64, P < 0.001). Interestingly, the popularity of this search term quickly plummeted after the epidemic peaked on week 40. For visualization purposes, the curve of the weekly number of new cases is square root transformed while the Google search trends indicate how often the terms “Ebola” and “Exponential Ebola” are searched for relative to the total number of searches (scale ranges from 0 to 100). The weekly series start with the first week in January 2014

Fig. 3

Representative time series of the cumulative number of weekly Ebola cases at the district level in Guinea, Sierra Leone, and Liberia. The district-level epidemics are spatially asynchronous and display an early growth phase that is more consistent with polynomial, rather than exponential, growth dynamics. The first week in the series ends on January 5, 2014

Fig. 4

Cumulative curves of four past Ebola outbreaks in Congo (1976, 1995, 2014) [–49] and Uganda (2000) [50]. These curves display rapid saturation in case growth within the first 3–4 generations of disease transmission, consistent with early sub-exponential growth dynamics

Fig. 5

Mean of the cumulative number of cases for the most affected districts of Liberia (as predicted by an agent-based model in Liberia [30]); patterns are consistent with sub-exponential growth dynamics

Fig. 6

Forecasting early epidemic growth phase data featuring sub-exponential growth dynamics using a classic exponential growth model (left) and the generalized growth model (right). The shaded region corresponds to the model calibration period and the non-shaded area corresponds to the forecasting period. Circles correspond to the case-series data. The blue curves correspond to the ensemble of epidemic forecasts. The red solid and dashed lines correspond to the median and interquartile range computed from the ensemble of forecasts, respectively. This figure illustrates how extrapolations of epidemic impact from the early growth trend in case incidence of an epidemic are subject to both model and data uncertainty. Transmission models calibrated using a few data points of the early phase of an infectious disease outbreak assuming exponential growth epidemic dynamics, such as the widely used SIR-type compartmental models, are unable to predict anything other than an exponentially growing epidemic in the absence of susceptible depletion, interventions or behavior changes, leading to great overestimation of cumulative case burden. More flexible transmission models, such as the generalized growth model, capture a wider range of epidemic growth profiles, ranging from sub-exponential to exponential growth dynamics. Please note the figures are on a different scale